US researchers have pointed to the potential risk of adversarial attacks on medical machine-learning systems. The adversarial examples that could be used are small changes to input data, that would be undetectable to humans, but would lead to drastically different output. This is potentially a huge problem for machine-learning-enabled diagnostic systems, medical billing, and insurance claims processing.
The researchers have called for an interdisciplinary approach to machine-learning and AI policy-making that would engage with medical, technical, legal, and ethical experts in healthcare.
#Health #AI #AdversarialExamples #AdversarialAttacks #Policy